Alpine Weather Forecasting Neil Stringfellow CSCS Swiss National Supercomputing Centre

CSCS – Swiss National Supercomputing Centre  National Supercomputing Centre since 1992  Provides compute facilities and scientific support to Swiss research community –Federal High Schools, Federal research institutes, Universities and University of Applied Sciences  Switzerland is currently planning its national strategy in HPC  CSCS also provides facilities to MeteoSwiss for operational weather forecasting HPC User Forum – Tucson - 9th Sept 2008

CSCS User Base  Scientists drawn from a large number of disciplines  Climate research is a major research field HPC User Forum – Tucson - 9th Sept 2008

Climate Modelling at CSCS  One of CSCS “ALPS” projects awarded to model hydrological cycle in Alpine Environment  Various software packages are run at CSCS –Echam5 & Echam5-HAM (Atmosphere & Aerosol) –CCSM & CSM with Carbon Cycle –COSMO climate model (regional and local)  No non-coupled ocean modelling HPC User Forum – Tucson - 9th Sept 2008

Economic Importance of Climate Modelling  Tourism –Important to know long-term effects for planning where to locate ski resorts  Agriculture –Swiss agriculture is expected to benefit from modest temperature increases (up to 2°C)  Electricity generation –Hydro power requires precipitation –Nuclear power plants require cooling HPC User Forum – Tucson - 9th Sept 2008

Water and Electricity Generation  Swiss electricity generation is carbon neutral! –Approx 60% from hydroelectric power plants –Most of the rest is Nuclear  Need to know precipitation levels for electricity generation  Cooling of nuclear power plants relies on water, and the temperature of that water –During the 2003 summer heatwave, electricity production from nuclear was reduced by 25% HPC User Forum – Tucson - 9th Sept 2008

Future Climate Scenarios  Current prediction is for higher temperatures and lower precipitation  Glacial melt will increase in near future but water available for hydro-generation will reduce from present levels by 2050  Warmer water will reduce cooling capacity for nuclear reactors  There is a need for research, and in particular numerical simulation HPC User Forum – Tucson - 9th Sept 2008

MeteoSwiss and CSCS  MeteoSwiss is the Swiss federal weather office  MeteoSwiss run operational weather forecast model at CSCS  MeteoSwiss runs the COSMO model from the COSMO consortium –This is a local (not global) model  CSCS provides compute resources and technical and scientific services HPC User Forum – Tucson - 9th Sept 2008

High Resolution Forecasting  European Windstorms Lothar and Martin caused destruction and loss of life in 1999  Not detected by national weather services  Demands for improved forecasting  Additional requirements for accurate forecasting from Nuclear Power Plant operators HPC User Forum – Tucson - 9th Sept 2008 Destruction in black forest due to Windstorm Lothar

European Windstorms - background  Windstorms occur in Winter, typically December to February –Sometimes called “Winter storms” or Orkan  Naming system similar to hurricanes –Names issued by Free University of Berlin  Actually all high and low pressures are named  Historically have caused major loss of life –Mainly due to dyke breaches in Netherlands  Occasionally missed by national weather services –1987 Storm in United Kingdom –“Lothar” in 1990 by Germany (and others inc. Switzerland) HPC User Forum – Tucson - 9th Sept 2008

Features of European Windstorms  Don’t dissipate quickly over land –They sometimes intensify over land  Often occur in clusters of 2 or more –Daria & Herta (Jan 1990) –Vivan & Wiebke (Feb 1990) –Désirée, Esther, Fanny, Hetty (Jan 1998) –Lothar & Martin (Dec 1999)  Wind speeds, insurance losses and fatalities are similar to U.S. hurricanes –No massive loss of life in modern times to compare with Hurricanes Jeanne and Katrina HPC User Forum – Tucson - 9th Sept 2008

Swiss Topography  High mountains and deep valleys lead to extreme winds during storms –225 km/h on Aetsch Glacier for Kyrill –285 km/h at Jungfraujoch for Wiebke HPC User Forum – Tucson - 9th Sept 2008

Insurance Losses  European Windstorms are the second highest cause of insurance losses –Highest losses are caused by U.S. Hurricanes  Average annual loss is around $2 Billion  5 of top 20 biggest ever insurance losses are due to European Windstorms HPC User Forum – Tucson - 9th Sept 2008

Losses of Big Storms Position in top 40 all time losses Wind StormYearLoss in US $Billion Fatalities 11Daria Lothar * Kyrill * “1987 Storm” Vivian * Martin * Anatol HPC User Forum – Tucson - 9th Sept 2008 * affected Switzerland Combined Lothar/Martin (25 th & 27 th Dec. 1999) would be 8 th largest loss Source: Swiss Re

Lothar/Martin – December 1999  Storm Lothar crossed France, Germany and Switzerland on 24 th & 25 th December 1999  Storm Martin followed a similar path on 26 th & 27 th December  Many fatalities, billions of dollars of damage  Not predicted by National Weather Services HPC User Forum – Tucson - 9th Sept 2008

Advances in Prediction  Study of prediction of Lothar/Martin (Walser et. al) looked at 3 aspects –Moist Singular Vectors  Different approach to calculate initial perturbations for ensemble forecasts –Increased Resolution –Ensembles  Showed great potential for improved forecasts HPC User Forum – Tucson - 9th Sept 2008

opr SVs,  x~80 km Forecast storm Lothar: max. wind gusts t+(42-66) (1) Configuration: opr SVs, 80 km opr SVs, 10 km, 80 km topo moist SVs, 10 km, 80 km topo moist SVs, 10 km moist SVs, 10 km, 10 members

Configuration: opr SVs, 80 km opr SVs, 10 km, 80 km topo moist SVs, 10 km, 80 km topo moist SVs, 10 km moist SVs, 10 km, 10 members opr SVs,  x~10 km,  x topography ~ 80 km Forecast storm Lothar: max. wind gusts t+(42-66) (2)

Configuration: opr SVs, 80 km opr SVs, 10 km, 80 km topo moist SVs, 10 km, 80 km topo moist SVs, 10 km moist SVs, 10 km, 10 members moist SVs,  x~10 km,  x topography ~ 80 km Forecast storm Lothar: max. wind gusts t+(42-66) (3 )

Configuration: opr SVs, 80 km opr SVs, 10 km, 80 km topo moist SVs, 10 km, 80 km topo moist SVs, 10 km moist SVs, 10 km, 10 members moist SVs,  x~10 km Forecast storm Lothar: max. wind gusts t+(42-66) (4)

Configuration: opr SVs, 80 km opr SVs, 10 km, 80 km topo moist SVs, 10 km, 80 km topo moist SVs, 10 km moist SVs, 10 km, 10 members moist SVs,  x~10 km, 10 members Forecast storm Lothar: max. wind gusts t+(42-66) (5)

Going from 80km to 10km HPC User Forum – Tucson - 9th Sept 2008 ECMWF EPS (80 km) COSMO-LEPS (10 km)

Current Situation of MeteoSwiss  Forecast runs on a 896 core Cray XT4  Runs 8 times per day for ~ 30 mins HPC User Forum – Tucson - 9th Sept 2008

Need for High Resolution  The forecast simulation resolves Switzerland using a two-grid refinement –coarse 6.6km spacing between grid points  385 x 325 grid, 60 atmospheric levels over Western Europe, 72 second time step with numerical leapfrog scheme –fine simulation uses 2.2km spacing  520 x 350 grid, 60 atmospheric levels over “Alpine Arc”, 20 second time step with Runge-Kutta numerical scheme  Many features in Switzerland were not resolved at the older 7km resolution –Few valleys are resolved at this resolution

Resolution change 6.6km to 2.2km HPC User Forum – Tucson - 9th Sept 2008 COSMO-7 (6.6 km)COSMO-2 (2.2 km)

HPC User Forum – Tucson - 9th Sept 2008 Example - Magadino Plain  Magadino Plain is the lowest part of Switzerland –Lowest point is on shore of Lago Maggiore  Plane is surrounded by mountains  At 6.6km res n it resolves to be a 1km high plateau  At 2.2km res n it has a valley floor at 200m height

Parameterisation v Direct Simulation  At low resolution many features cannot be directly modelled - have to be parameterised  Higher resolutions allow more physics  6.6km -> 2.2km deep convection is computed explicitly  Higher resolution also allows modelling of valley winds HPC User Forum – Tucson - 9th Sept 2008

Full Suite  7 components –Interpolation, assimilation and 24 hour forecast on coarse grid –Interpolation and assimilation on fine grid –Interpolation and 24 hour forecast on fine grid  All components have to complete in ~20 minutes –To allow for data post-processing to complete within 30 minutes of start  Suite runs every 3 hours  Twice per day a 72 hour coarse grid forecast is added HPC User Forum – Tucson - 9th Sept 2008

Model Heirarchy HPC User Forum – Tucson - 9th Sept 2008 ECMWF IFS ECMWF IFS (global) 25km, 91 levels 2 x 240h per day + 2 x 78h per day COSMO-7 (regional) 6.6km, 60 levels 2 x 72h per day COSMO-7 COSMO-2 (local) 2.2km, 60 levels 8 x 24h per day COSMO-2

Full Suite Timeline HPC User Forum – Tucson - 9th Sept COSMO-7 COSMO-2 Time UTC.. +72h Full Suite Forecast

Example of Improvement - Wind  South of Zurich Lake  Wind field at 6.6km and 2.2km resolution  Features only resolved at high resolution HPC User Forum – Tucson - 9th Sept 2008

Other Extreme Events in Switzerland  Summer Flooding –Summer floods over central Europe in 2005 –38 th largest insurance loss (Swiss Re)  Summer Heatwaves –European heatwave of 2003 responsible for 35,000 deaths  8 th largest number of deaths from natural catastrophe  Others, e.g. hailstorms halted Tour de Suisse in 2007 HPC User Forum – Tucson - 9th Sept 2008

HPC Issues in Climate/Weather  What is typical high-end Climate HPC work?  Future Modelling in Climate/Weather –Higher resolution –More physics –Ensembles  Very complex and large codes –Not likely to be an early adopter or new languages –No compact kernel for accelerators HPC User Forum – Tucson - 9th Sept 2008

I/O Rate and Storage  Many codes use proprietary formats –Grib format in European codes  No widespread adoption of parallel I/O –often I/O is done on one or a few processes  Increasing amounts of data being generated –reluctance to delete data –two-thirds of CSCS archive is used for Climate and Weather data HPC User Forum – Tucson - 9th Sept 2008

Acknowledgements  Great many thanks go to Andre Walser and Daniel Leuenberger of MeteoSwiss for providing slides and answering questions HPC User Forum – Tucson - 9th Sept 2008